An 80-year-old problem in discrete geometry, first posed by legendary Hungarian mathematician Paul Erdős, has been solved in a way that contradicts his original conjecture, according to findings announced by OpenAI and independently verified by external mathematicians. The result, reported in coverage referenced by Nature, suggests that an AI system may have surpassed human expectations in tackling one of mathematics’ enduring challenges using a single prompt.
The problem at the centre of the discovery is known as the unit-distance problem, which concerns how many pairs of points in a plane can be arranged so that they share the same distance. Erdős believed he had effectively set an upper limit on this configuration in 1946, proposing that no arrangement could significantly outperform his construction as the number of points grows. For decades, his intuition shaped the field and stood largely unchallenged.
That assumption has now been overturned, at least according to OpenAI’s announcement on 20 May. The company claims its chatbot system was able to construct a counterexample using techniques from algebraic number theory, allowing it to identify specific coordinate systems derived from carefully chosen equations. These mathematical structures enabled a configuration of points that exceeds the bounds Erdős believed to be optimal.
The result has been independently checked by mathematicians not affiliated with OpenAI, adding weight to the claim that the conjecture does not hold. While the company has not released the full details of the 125-page reasoning process or disclosed the exact model used, researchers familiar with the work say the output appears to be internally consistent and mathematically rigorous.
Reactions from the mathematical community have been intense. Tom Trotter of the Georgia Institute of Technology, who co-authored work with Erdős, said he believed the late mathematician would have been “raving” about the development if he were alive today. Others have expressed similar astonishment at the idea that a machine system has produced a result of such conceptual depth in a field known for its resistance to automation.
Sebastien Bubeck, a mathematician at OpenAI, described the breakthrough as potentially the first instance of an AI system autonomously generating a meaningful result in research mathematics. Speaking about the process, he emphasized that the model was not guided step-by-step toward a solution but instead responded to a single open-ended prompt asking whether Erdős’s conjecture might be true or false.
According to Bubeck, the system produced a long, continuous chain of reasoning rather than iterating through multiple corrected attempts. In contrast to earlier AI approaches that rely on repeated refinement and human-guided prompting, this model reportedly interpreted the problem in a single pass and generated a complete mathematical argument. Nature’s reporting on the development highlights how unusual this is compared with previous AI-assisted proofs, which typically require extensive human intervention.
The company describes the underlying model as a general-purpose reasoning system rather than a specialized mathematical tool. It was not explicitly designed for theorem proving, yet it was able to apply advanced methods from algebraic number theory to a domain problem in discrete geometry. This cross-domain capability has been highlighted by researchers as a key indicator of shifting AI capabilities, where systems are no longer confined to narrow areas of expertise.
Mathematicians who reviewed the work, including Daniel Litt of the University of Toronto, have noted that while the full reasoning has not been released, the independently verified portions appear valid and conceptually interesting. Litt remarked, as reported in Nature, that this may represent the first instance of an AI-generated result that is not only correct but also inherently meaningful within mathematical research.
The achievement also highlights a broader shift in how AI systems are being used in scientific discovery. Unlike traditional “orchestration” methods, where large language models are repeatedly prompted and corrected in iterative cycles, this approach suggests a more direct form of problem solving. The model’s output reportedly remains stable regardless of how the initial question is phrased, implying a robustness that resembles human mathematical reasoning more closely than previous systems.
OpenAI researchers argue that this ability to generalize across domains is what makes the result significant. Instead of relying on narrow training for mathematical tasks, the model appears to draw from a broad base of knowledge and reasoning strategies. This allows it to navigate unfamiliar areas of mathematics without explicit domain-specific instruction, a capability that has long been considered out of reach for machine systems.
Still, the lack of transparency surrounding the full proof and model architecture has left parts of the academic community cautious. Without full access to the 125-page reasoning chain, independent verification remains limited, even if initial checks support the validity of the result. This tension between openness and proprietary AI development continues to shape debates about how such discoveries should be evaluated.
Despite these uncertainties, the excitement in the field is unmistakable. Some researchers, including those at OpenAI, have suggested that this may be one of the clearest examples yet of an AI system contributing independently to fundamental mathematical research. Others, cited in Nature’s reporting, have described it as a breakthrough that challenges assumptions about the pace at which artificial intelligence can advance scientific knowledge.
The implications extend beyond a single geometric problem. If AI systems can independently generate new mathematical insights, they may eventually play a central role in fields that rely on abstract reasoning and proof construction. However, researchers also caution that mathematics is uniquely suited to formal verification, whereas other scientific domains require experimental validation that AI systems cannot yet perform alone.
For now, the resolution of Erdős’s 80-year challenge stands as a symbolic moment. A problem once thought to define the limits of human intuition in geometry has been revisited—and potentially overturned—by an artificial intelligence system operating from a single prompt. Whether this marks a turning point in mathematical discovery or an exceptional one-off result remains an open question, but as Nature’s coverage suggests, the boundary between human and machine reasoning in mathematics is beginning to blur in ways few anticipated.
